vLLM: Resource-Exhaustion (DoS) through Malicious Jinja Template in OpenAI-Compatible Server
Summary
A resource-exhaustion (denial-of-service) vulnerability exists in multiple endpoints of the OpenAI-Compatible Server due to the ability to specify Jinja templates via the chat_template and chat_template_kwargs parameters. If an attacker can supply these parameters to the API, they can cause a service outage by exhausting CPU and/or memory resources.
Details
When using an LLM as a chat model, the conversation history must be rendered into a text input for the model. In hf/transformer, this rendering is performed using a Jinja template. The OpenAI-Compatible Server launched by vllm serve exposes a chat_template parameter that lets users specify that template. In addition, the server accepts a chat_template_kwargs parameter to pass extra keyword arguments to the rendering function.
Because Jinja templates support programming-language-like constructs (loops, nested iterations, etc.), a crafted template can consume extremely large amounts of CPU and memory and thereby trigger a denial-of-service condition.
Importantly, simply forbidding the chat_template parameter does not fully mitigate the issue. The implementation constructs a dictionary of keyword arguments for apply_hf_chat_template and then updates that dictionary with the user-supplied chat_template_kwargs via dict.update. Since dict.update can overwrite existing keys, an attacker can place a chat_template key inside chat_template_kwargs to replace the template that will be used by apply_hf_chat_template.
# vllm/entrypoints/openai/serving_engine.py#L794-L816
_chat_template_kwargs: dict[str, Any] = dict(
chat_template=chat_template,
add_generation_prompt=add_generation_prompt,
continue_final_message=continue_final_message,
tools=tool_dicts,
documents=documents,
)
_chat_template_kwargs.update(chat_template_kwargs or {})request_prompt: Union[str, list[int]]
if isinstance(tokenizer, MistralTokenizer):
...
else:
request_prompt = apply_hf_chat_template(
tokenizer=tokenizer,
conversation=conversation,
model_config=model_config,
**_chat_template_kwargs,
)
Impact
If an OpenAI-Compatible Server exposes endpoints that accept chat_template or chat_template_kwargs from untrusted clients, an attacker can submit a malicious Jinja template (directly or by overriding chat_template inside chat_template_kwargs) that consumes excessive CPU and/or memory. This can result in a resource-exhaustion denial-of-service that renders the server unresponsive to legitimate requests.
Fixes
* https://github.com/vllm-project/vllm/pull/25794
What the EG score is made of (scorer v4.0: the three questions the 0–10 score mixes, published apart · methodology)
The EG score (6.5): severity, escalated by confirmed exploitation. This CVE's 6.5 is its impact severity (EG Severity), not escalated.
- EG Severity
- 6.5 MEDIUMCVSS-B, impact only: no KEV listing, EPSS or EG-KEV tier raises it.GitHub Security Advisory CVSS
- EG Likelihood
- Unknown (no EPSS on record)Unknown, not low: FIRST has published no EPSS probability for this CVE.
- Confirmed exploitation
- None on record (CISA KEV, EG-KEV Tier 1)Reported apart from the likelihood: an observed fact, not a probability.
- EG Priority
- SSVC Track at medium mission impactTrack at low or medium mission impact; Track* at high (mission-essential systems).CISA BOD 26-04: 14 days if the affected system is publicly exposed; Fix on system upgrade if it is not.Each input and its source, and whom BOD 26-04 binds: the SSVC and BOD 26-04 blocks below.
SSVC (CISA's decision table, applied by EchelonGraph)Track at low or medium mission impact; Track* at high (mission-essential systems).
A fix is available. Apply it within your standard update timelines.
Exploitation none (no CISA value, KEV listing or exploit code on record) · Automatable no (CISA's BOD 26-04 default) · Technical impact total (CISA's BOD 26-04 default). EchelonGraph holds no SSVC assessment of this CVE from CISA: none of these inputs is from Vulnrichment. EchelonGraph applied CISA's decision table to them; CISA publishes SSVC inputs and the table, not a decision for each CVE. Mission impact is CISA's Mission & Well-being decision point, and only you can judge it. CISA's table makes it high when Mission Prevalence is Essential — the vulnerable component "directly provides capabilities that constitute at least one MEF for at least one entity" (MEF: mission essential function) — or when Public Well-Being Impact is Irreversible: "multiple fatalities are likely", the cyber-physical system "is likely lost or destroyed", "extreme or serious externalities" are imposed on other parties, or social systems such as elections or the financial grid "are destabilized and potentially collapse". CISA's decision table
BOD 26-04 (CISA's remediation timeline, Table 1)14 days if the affected system is publicly exposed; Fix on system upgrade if it is not.
The score's inputs: In CISA KEV no (CISA's KEV catalog) · Automatable no (CISA's BOD 26-04 default) · Technical impact total (CISA's BOD 26-04 default). Where an input reads “CISA's BOD 26-04 default”, no one has published it for this CVE, and CISA's implementation guidance says: “If “Automatable” and “Technical Impact” information items about a CVE ID are not available and the CVE ID is not in the KEV Catalog, CISA will treat its values as “no” and “total,” respectively, until CISA or a third party provides metadata about the CVE ID. CISA will always provide the Vulnrichment data for a CVE ID listed in the KEV Catalog.” Whether your system is publicly exposed is yours to answer. BOD 26-04 binds US Federal Civilian Executive Branch agencies; for anyone else it is CISA's published timeline, a reference rather than an obligation. The directive · how EchelonGraph applies it